Large language models preserve a geometric substrate in value representations that correlates with uncertainty and matches patterns from small models performing exact Bayesian inference.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Geometric Scaling of Bayesian Inference in LLMs
Large language models preserve a geometric substrate in value representations that correlates with uncertainty and matches patterns from small models performing exact Bayesian inference.